Adarsh Singh

AI Researcher

New Delhi, Delhi, India3 yrs experience

AI EnabledAI ML Practitioner

Key Highlights

Expert in building reliable AI systems under real-world constraints.
Proven track record in optimizing production NLP pipelines.
Skilled in developing agentic systems with reduced hallucination rates.

Stackforce AI infers this person is a Data Processing and AI Solutions Expert specializing in production-grade systems.

Contact

Skills

Core Skills

Large Language Models (llm)Natural Language Processing (nlp)Distributed ComputingData Privacy AwarenessRetrieval-augmented Generation (rag)Multi-agent SystemsObservabilityGenerative Ai

Other Skills

Graph AlgorithmsTransformer Re-rankingToken-budgeted context compressionScrapyRedisCeleryLlama-3GPT-4MilvusNeo4jLangGraphDockerOpenTelemetryState-managementReal-time response buffering

About

I build AI systems that have to work under real-world constraints. My work focuses on the gap between model capability and production reliability—across hybrid retrieval (vector + graph + re-ranking), agentic state machines, and inference-time optimization. I’m particularly interested in systems that can reason over messy data, know when to stop, and remain observable and debuggable at scale. Currently building production OSINT pipelines (20k+ multilingual docs/day) with sub-100ms latency and cost-efficient model routing. Interested in reliable RAG infrastructure, agentic systems, and LLM pipelines that behave like real software.

Experience

3 yrs

Total Experience

1 yr 9 mos

Average Tenure

1 yr 3 mos

Current Experience

Crimson energy experts pvt ltd

2 roles

LLM Engineer

Feb 2025 – Present · 1 yr 3 mos

Built a production OSINT + threat intelligence platform for a defence
client — owned architecture and implementation end-to-end.
▸ GraphRAG retrieval fusing vector semantic search with knowledge graph
multi-hop traversal, transformer re-ranking, and token-budgeted context
compression for real-time queries.
▸ Multi-provider LLM routing across cloud and self-hosted models —
profile-mapped per task, with offline fallback. 75% API cost reduction.
▸ Sensor-driven orchestration pipeline processing high-volume multilingual
news at scale, with batched NLP (classify → embed → NER → graph build).
▸ Stateful LangGraph multi-agent system with evidence buffering and
confidence thresholding — 26% hallucination reduction vs zero-shot baseline.
▸ Full observability stack: metrics, structured logging, GPU monitoring,
encrypted secrets management.

Large Language Models (LLM)Natural Language Processing (NLP)

LLM Engineer

Feb 2025 – Present · 1 yr 3 mos

Distributed Ingestion Architecture: Re-architected a monolithic scraping pipeline into a microservices cluster using Scrapy, Redis, and Celery. Implemented distributed locking (Redlock) to coordinate 50+ concurrent spiders, scaling throughput to 20,000+ articles/day while preventing IP bans.
Hybrid Inference Strategy: Designed a Model Routing system that processes high-volume extraction tasks via self-hosted Llama-3 (vLLM) and routes complex reasoning to GPT-4. This tiered approach reduced API costs by 75% while ensuring data privacy for sensitive raw ingest.
GraphRAG Optimization: Built a production-grade RAG pipeline combining Milvus (HNSW) and Neo4j (Knowledge Graph). Utilized local BGE-M3 embeddings for re-ranking, optimizing recall vs. latency trade-offs to achieve <100ms retrieval for chat queries.
Agentic Reliability: Designed a Stateful Multi-Agent System (LangGraph) for geopolitical risk analysis with memory state. Implemented recursive verification loops that reduced hallucination rates by 26% compared to zero-shot baselines.
Observability: Deployed the stack on Docker, instrumenting OpenTelemetry traces to visualize the hand-off latency between local models and external APIs, reducing system Mean Time To Resolution (MTTR) by 40%.

Hd softwares

2 roles

AI Developer

May 2023 – Feb 2025 · 1 yr 9 mos · Noida, Uttar Pradesh, India

Production Agentic Systems: Developed a low-latency AI Receptionist that automated 70% of Tier-1 support calls, focusing on state-management and real-time response buffering.
Domain Adaptation (LoRA): Fine-tuned CodeLlama-7B for internal proprietary codebases, reducing developer boilerplate time by 35%.
E2E Deployment: Managed the full lifecycle of LLM deployment, from prompt engineering and evaluation frameworks to monitoring performance in production environments.

Generative AILarge Language Models (LLM)

AI Developer Intern

Feb 2023 – May 2023 · 3 mos · Noida, Uttar Pradesh, India

During my internship at HD Softwares, I had the opportunity to dive into the basics of AI and machine learning, working closely with experienced developers on real-world projects. My role involved:
Model Training and Evaluation: I helped train and evaluate machine learning models, which were used in developing AI-driven solutions for the company’s internal projects.
Data Preprocessing: I worked on preparing datasets by performing data preprocessing and feature engineering, ensuring the data was in the best shape for model training.
Research and Prototyping: I spent time researching new AI technologies and creating simple prototypes to explore how they could be applied to ongoing projects.
Project Support: I provided support across various AI projects, collaborating with senior developers to implement and improve algorithms.
During this internship, I learned the essentials of AI development, including gaining hands-on experience with the PyTorch framework. This experience laid the groundwork for my current role, where I now take on more advanced challenges in AI development.

PyTorchNatural Language Processing (NLP)

Zummit infolabs

Data Science Intern

Jul 2022 – Dec 2022 · 5 mos · Bengaluru, Karnataka, India · Remote

As a Data Science Intern at Zummit Infolabs, I honed my skills in machine learning by developing innovative models tailored for real-world applications. My projects included creating a yawn detection model to monitor alertness, a news classification system to filter and categorize content, and a speaker recognition model to enhance security measures. This role enabled me to leverage analytical tools and techniques to solve complex problems and deliver data-driven solutions.

Python (Programming Language)Exploratory Data Analysis